A recorded run

See it work, no keys required.

The live demo gates real runs behind an operator key so a public URL cannot run up an API bill. This is a real run, captured end to end - so you can see the whole thing without one: a live trace, a human-in-the-loop approval, a cited report, and the audit trail straight from the database.

Run

Give an agent a real task

A research agent, configured with web search and source-fetch tools on Haiku for cheap retrieval. The task: identify the current events and tensions between Anthropic and the US government. It searches, then fetches each primary source before it will cite it - it cannot cite what it has not read.

Approve

Nothing ships without a human

The agent finished its research and wants to publish its brief - but it pauses. Publishing is a high-risk action, so it stops for a human decision, with the full proposed brief and its rationale in view. Approve or reject. Nothing is published autonomously.

Report

A cited report, after sign-off

Once approved, the brief publishes: an executive summary and key developments, each tied to a primary source the agent actually fetched. 39 steps, ~93k tokens, fully traced.

The audit trail

Trust, but verify - in the database.

Approvals and reasoning are not UI theater. They are rows you can query.

Audit

Every run, step, and decision is persisted

Runs, steps, tool calls, approvals, findings, and briefs all land in a relational schema. The agent's entire trace - including its private reasoning - is stored, not just the final answer.

Audit

Query what the agent was thinking

Because the reasoning is stored, you can query it directly. This pulls the agent's thinking steps for the run straight from the database - the answer to 'why did it do that?' is a SQL query away.

Audit

Prove a human approved it

The strongest audit question - 'did a person actually approve this action?' - is answerable. This joins the run's publish_brief tool call to its approval record: the exact action, the brief content, and the decision (approved). Defensible governance, not a screenshot of a chat.